Skip to content
be-human.ai
← Back to the journal

Testing & reliability

Agent Drift Detection for Superintelligent Agents: Quiet Quality Loss

Catch gradual AI agent quality loss before customers do. A weekly sample-and-score routine, a frozen baseline and review thresholds for business teams.

Be Human typographic cover with the title: Agent Drift Detection for Superintelligent Agents: Quiet Quality Loss
Conceptual illustration created with AI · Be Human

Agent drift detection is the practice of re-scoring an AI agent's finished work on a fixed schedule, against a standard that does not change, so that gradual quality loss shows up as a trend in your own data instead of as complaints from customers. The most dependable starting method for business teams is a weekly sample-and-score routine: pull a small random set of completed runs, score each one against a short rubric, and compare the results with a baseline recorded while the agent was performing well. The routine is built to catch small movements early, because drift rarely produces a single obvious failure.

Key takeaways

  • Drift is a slow change in output quality, and it rarely triggers an error, so uptime and error dashboards can look healthy while it happens.
  • A weekly random sample of real completed work, scored on a fixed rubric, is enough to start and does not require a data science team.
  • Keep a frozen baseline for each task type so that every week is compared with the same standard.
  • Decide before the first review which score movement opens a review, who receives it, and which actions the agent must pause in the meantime.
  • Look upstream as well as at the agent, because knowledge sources, tool formats and model versions often change output before anyone edits the agent.

What drift looks like when nothing has broken

Drift is the gap that opens between how an agent performed when it was set up and how it performs now, with no single event to blame. An agent that once cited the current return policy begins citing an older one. Replies grow more generic, a required field is skipped now and then and then more often, and escalations to a person become rarer. That last change looks like efficiency until the cases that should have been escalated start returning as complaints.

Three conditions make drift hard to see. Each run passes technical checks, the change happens in small steps, and the people who notice first are usually customers. By the time a complaint arrives, the pattern has often been present for several weeks. A sample-and-score routine turns a vague feeling that replies seem worse into a measured line on a chart that a named person is responsible for reading. For a wider view of the symptoms, see AI agent drift: how to spot quality decline over time.

Set the baseline before you measure any change

You cannot detect drift without a reference point. Choose three to five task types the agent handles most often, such as answering policy questions, drafting follow-up messages or classifying incoming requests. For each one, pull runs from a period when the team reviewed the outputs and accepted them, then score those runs with the rubric you will use from now on. The average and the spread of those scores become the baseline band.

Write the rubric as four or five criteria, each with a short definition and three possible scores: pass, partial and fail. Typical criteria are factual accuracy against current sources, use of the required format, tone that matches the brand, completeness of required fields and correct escalation. Keep each definition concrete enough that two reviewers would give the same score to the same reply. Public guidance on evaluating agents, such as Anthropic's notes on demystifying evals for AI agents, is a useful reference when you draft these definitions.

Freeze the baseline and store it with a version label. When you improve the agent, create a new baseline only after you have re-scored the old cases, so that a change in the standard is never mistaken for a change in the agent. A small set of approved cases that you rerun after each change is a golden test set, and it is the natural companion to this routine. See golden test sets for superintelligent agents for how to build one.

The weekly sample-and-score routine, step by step

Assign a fixed day and a named owner. Once the shared sheet exists, the routine takes less time than most teams expect. Run it in this order:

  • Export the completed runs for each task type from the past seven days, including the input, the output, the sources used and any tool calls.
  • Draw a random sample from each export and record the selection method so the sample can be reproduced later.
  • Remove customer names, contact details and account numbers before a reviewer opens an item, and keep the run identifier so the full record can be found when needed.
  • Score each sampled item against the rubric without looking at last week's scores, so reviewers are not anchored by earlier results.
  • Write one line of note for every partial or fail, naming the criterion and what was wrong.
  • Log the scores in one shared table with the week, task type, reviewer and run identifier.
  • Compare this week's average and fail count for each criterion with the baseline band.
  • Read the low-scoring items together for about ten minutes and name the pattern. Do not stop at the individual mistake.

The figure below shows the idea behind the sample: each batch is checked in the same way, so that differences between batches are visible rather than assumed.

How large the sample should be and who scores it

The right sample size depends on volume and on what a bad answer costs. A task type that touches money, legal commitments or customer records deserves a larger sample than an internal summary task. Start with enough items per task type that one unusual reply cannot move the average on its own, then look at how stable the results are and adjust. For a fuller method, see how many AI agent outputs to review. The NIST AI Risk Management Framework is also a useful reference for deciding how much oversight a given task deserves.

Reviewers matter as much as sample size. Have at least two people score a calibration subset each month, each working independently. Where they disagree, revise the rubric definition rather than averaging the disagreement away, because disagreement usually means the criterion is unclear. If an automated judge scores some items, compare its results with human scores first and keep checking that comparison, since a judge can drift as well.

A decision guide for score movements

Before the first review, write down what each movement means. A simple guide:

  • A week inside the baseline band: record it and keep the normal sample size.
  • A single week below the band on one criterion: double the sample for that task type the following week and note the change.
  • Two consecutive weeks below the band on the same criterion: open a review with the task owner and name the likely cause, such as a source change or a recent prompt edit.
  • Any fail on an action that moves money, changes a booking or sends a commitment to a customer: pause that action type the same day and review the runs before it resumes.
  • A recovery after a known change: record the cause in the change log so the next reader knows what happened.

Name an owner for each row. The routine protects you only when each row leads to an action that someone is allowed to take.

Where drift starts before the agent changes

Drift often begins outside the agent. A knowledge source is updated, but the search index still returns an older document. A tool starts returning a renamed field, and the agent quietly ignores it. A model version changes behavior on a subtle instruction. The mix of incoming requests shifts toward a case the agent handles less well. The agent's own instructions may also grow through small edits that each looked harmless on their own.

Illustrative scenario: a support agent answers questions about a return window. Over three weekly reviews, the criterion for using the current policy moves from pass to partial on a growing number of items, while tone scores stay steady. Reviewers trace the problem to a policy document that was replaced in the source folder, while retrieval still returns the older copy. The team restores the current source, adds a freshness check on that document, reruns the golden test set and records the change. No customer had complained during the first two weeks, which is why the trend mattered more than the complaints would have.

Starting the routine in your team

Pick the one task type where a quiet decline would cost you the most, write its rubric, freeze a baseline and run the sample-and-score routine for four weeks before you judge its value. Be Human designs and operates this kind of oversight as part of managed operations, which can include monitoring, recurring evaluations, permission review and controlled improvements. In practice, the rubric, the baseline and the score log remain under the client's control, samples are drawn from the agent's own logged runs, and each change that moves a score is recorded and reviewed by a person before it reaches production. Be Human does not promise a particular score or outcome. The routine exists to make quality visible and to make the next decision easier to take. If you are not yet sure which of your processes should be the first to get this kind of oversight, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

How is drift detection different from error monitoring? Error monitoring catches runs that fail outright, such as timeouts, rejected tool calls or exceptions. Drift detection catches runs that succeed technically but gradually produce weaker answers, which error dashboards will show as healthy. The two work best together, with error alerts for sudden breaks and a weekly sample for slow change.

How many outputs should I sample each week? Begin with enough items per task type that one unusual reply cannot change the average on its own, and raise the sample for tasks that touch money, commitments or customer data. Keep the sample at a size your reviewers can score carefully, because a rushed review is worse than a smaller one. Adjust after the first month, based on how stable the scores have been.

Can an AI judge do the scoring instead of people? An automated judge can handle routine checks well, such as whether a required field is present, but it should be calibrated against human scores before anyone relies on it. Have people score a calibration subset each month and compare the results. Keep people on the criteria where judgment about accuracy or tone matters most.

What if we have no baseline yet? You can build one from work the team has already reviewed and accepted, even if that review was informal. Score those past outputs with the new rubric and treat the result as the first baseline band. Label it with its period and the reviewers involved, and do not treat it as a permanent standard until it has held up through a few weeks of use.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →